Unmanned aerial vehicles, known as UAVs, have evolved over the past two decades to sophisticated aircraft robots able to carry out surveillance, recognition, remote sensing and even attack missions. But there are not many alternatives of autonomous navigation systems for most of these aircraft which still require human intervention to navigate. Devices such as Global Positioning System (GPS) and inertial systems help calculate routes and locate the vehicle on a map among other possibilities, but do not offer solutions to unknown or uncertain circumstances. On the other hand, computer vision techniques have provided many possible applications for intelligent systems such as object recognition, robot localization and reconstruction of 3D maps. This paper explores the use of computer vision and pattern recognition techniques for UAV navigation, and proposes a set of visual features based on color and gradients orientation for image classification. To validate the proposed approach, a system was developed to evaluate the classification and matching of aerial images. The results achieve more than 95% of accuracy and confirm the viability of the selected algorithms and methods for the problem.

Due to the widespread acceptance of Synthetic Aperture Radar (SAR) imagery by the scientific community in recent years, SAR system design faces ever increasing demands for wide coverage of high resolution images. This represents a challenging task, since conventional monostatic SAR system are inherently subject to a compromise between the achievable resolution and the width of the imaged swath. In this context, multi-channel SAR systems arise as a promising alternative to overcome this limitation and achieve high-resolution wide-swath (HRWS) SAR imaging. Their operation requires the acquisition of more information about the scene through multiple channels and suitable digital beamforming techniques to adequately combine the output data. The innovative Multi-Channel Reconstruction Algorithm (MCRA) was recently introduced as a suitable alternative for the processing required by multi-channel SAR systems in azimuth. This thesis performs and evaluates on a demonstration of the capabilities of this algorithm employing real SAR multi-channel data acquired by DLR's new airborne system F-SAR. The combination of up to four different channels is performed, and analyzes of the algorithm performance follow, especially with regard to ambiguity suppression. The impact of channel imbalances in the residual ambiguity levels is considered and different channel balancing methods are assessed. Following the algorithm's success in yielding high quality reconstructed images given adequately balanced channels, focus is turned to balancing alternatives, with special interest in methods applicable to aliased data. Blind equalization techniques are employed to develop a version of the reconstruction algorithm with increased robustness to channel imbalances following a certain error model. Performance assessments of the alternative strategy are performed both for controlled imbalances following the design model and real unknown imbalances found on F-SAR data.